未验证 提交 6c80bb3c 编写于 作者: X Xin Pan 提交者: GitHub

Merge pull request #14786 from PaddlePaddle/revert-14782-revert-14398-imperative

Revert "Revert "Imperative""
add_subdirectory(memory)
add_subdirectory(platform)
add_subdirectory(framework)
add_subdirectory(imperative)
add_subdirectory(operators)
add_subdirectory(string)
add_subdirectory(recordio)
......
......@@ -16,7 +16,9 @@ limitations under the License. */
#include <string>
#include <vector>
#include "glog/logging.h"
#include "paddle/fluid/framework/var_type.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/platform/place.h"
namespace paddle {
namespace framework {
......@@ -53,5 +55,12 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
return tensor;
}
LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name) {
Variable* var = scope.FindVar(var_name);
PADDLE_ENFORCE(var, "%s no in scope", var_name);
PADDLE_ENFORCE(var->IsType<LoDTensor>(), "Only support lod tensor now.");
return *var->GetMutable<LoDTensor>();
}
} // namespace framework
} // namespace paddle
......@@ -27,5 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
size_t index);
LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name);
} // namespace framework
} // namespace paddle
......@@ -38,8 +38,7 @@ void CheckProgram(const ProgramDesc &program) {
switch (role_id) {
case _INT(OpRole::kForward):
if (visit.find(_INT(OpRole::kBackward)) != visit.end()) {
LOG(ERROR)
<< "Cannot add backward operator before forward operator %s."
LOG(ERROR) << "Cannot add backward operator before forward operator "
<< op->Type();
}
break;
......
cc_library(layer SRCS layer.cc DEPS proto_desc operator)
cc_library(tracer SRCS tracer.cc DEPS proto_desc)
cc_library(engine SRCS engine.cc)
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/imperative/engine.h"
#include <mutex> // NOLINT
#include <vector>
#include "glog/logging.h"
namespace paddle {
namespace imperative {
static std::once_flag init_engine;
static Engine* engine;
class DummyEngine : public Engine {
public:
void Enqueue(Runnable* runnable) override {
queued_runnables_.push_back(runnable);
}
size_t Size() const override { return queued_runnables_.size(); }
void Sync() override {
for (Runnable* l : queued_runnables_) {
LOG(INFO) << "running " << reinterpret_cast<void*>(l);
}
queued_runnables_.clear();
}
private:
std::vector<Runnable*> queued_runnables_;
};
Engine* GetEngine() {
std::call_once(init_engine, []() { engine = new DummyEngine(); });
return engine;
}
} // namespace imperative
} // namespace paddle
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <cstddef>
#include <cstdint>
namespace paddle {
namespace imperative {
struct Runnable {};
class Engine {
public:
virtual ~Engine() {}
virtual void Enqueue(Runnable* runnable) = 0;
virtual size_t Size() const = 0;
virtual void Sync() = 0;
};
Engine* GetEngine();
} // namespace imperative
} // namespace paddle
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/imperative/layer.h"
#include <deque>
#include <limits>
#include <map>
#include <random>
#include <utility>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/string/printf.h"
namespace paddle {
namespace imperative {
using framework::Variable;
void AddTo(Variable* src, Variable* dst) {
framework::LoDTensor* dst_tensor = dst->GetMutable<framework::LoDTensor>();
framework::LoDTensor* src_tensor = src->GetMutable<framework::LoDTensor>();
PADDLE_ENFORCE(dst_tensor->numel() == src_tensor->numel(), "%lld vs %lld",
dst_tensor->numel(), src_tensor->numel());
float* dst_data = dst_tensor->mutable_data<float>(platform::CPUPlace());
const float* src_data = src_tensor->data<float>();
for (size_t i = 0; i < src_tensor->numel(); ++i) {
dst_data[i] += src_data[i];
}
}
class Autograd {
public:
explicit Autograd(framework::Scope* scope) : scope_(scope) {}
void RunBackward(VarBase* var) {
PADDLE_ENFORCE(var->pre_op_->op_desc_);
// TODO(panyx0718): Only create for vars that "require_grad"
(*var->pre_op_->output_vars_)[var->pre_op_out_idx_]->grads_ = var->grads_;
std::deque<OpBase*> ready;
ready.push_back(var->pre_op_);
std::map<OpBase*, int> dep_counts = ComputeDepCounts(var->pre_op_);
while (!ready.empty()) {
OpBase* ready_op = ready.front();
ready.pop_front();
std::vector<Variable*> input_grads = ready_op->ApplyGrad(scope_);
for (size_t i = 0; i < input_grads.size(); ++i) {
if (!input_grads[i]) continue;
OpBase* pre_op = ready_op->pre_ops_->at(i);
if (!pre_op) continue;
dep_counts[pre_op] -= 1;
PADDLE_ENFORCE(dep_counts[pre_op] >= 0);
bool pre_op_ready = dep_counts[pre_op] == 0;
if (pre_op_ready) {
ready.push_back(pre_op);
}
}
}
}
private:
std::map<OpBase*, int> ComputeDepCounts(OpBase* op) {
std::map<OpBase*, int> ret;
std::deque<OpBase*> queue;
queue.push_back(op);
std::unordered_set<OpBase*> visited;
visited.insert(op);
while (!queue.empty()) {
OpBase* candidate = queue.front();
queue.pop_front();
for (OpBase* pre_op : *(candidate->pre_ops_)) {
if (!pre_op) continue;
if (visited.find(pre_op) == visited.end()) {
visited.insert(pre_op);
queue.push_back(pre_op);
}
ret[pre_op] += 1;
}
}
return ret;
}
framework::Scope* scope_;
};
framework::Variable* CreateVariable(const std::string& name,
const framework::DDim& dim, float val,
framework::Scope* scope,
bool random_name = true) {
std::string varname = name;
if (random_name) {
std::mt19937 rng;
rng.seed(std::random_device()());
std::uniform_int_distribution<std::mt19937::result_type> dist6(
1, std::numeric_limits<int>::max());
int id = dist6(rng);
varname = string::Sprintf("%s@%d", varname, id);
}
VLOG(3) << "creating var " << varname;
framework::Variable* var = scope->Var(varname);
framework::LoDTensor* tensor = var->GetMutable<framework::LoDTensor>();
float* data = tensor->mutable_data<float>(dim, platform::CPUPlace());
std::fill(data, data + tensor->numel(), val);
return var;
}
framework::LoDTensor& VarBase::Grad() {
VLOG(3) << "get var grad " << var_desc_->Name();
return *grads_->GetMutable<framework::LoDTensor>();
}
void VarBase::ApplyGrad(framework::Scope* scope, Variable* grad) {
VLOG(3) << "apply var grad " << var_desc_->Name() << " "
<< grad->Get<framework::LoDTensor>().data<float>()[0];
if (!grads_) {
grads_ =
CreateVariable(string::Sprintf("%s@IGrad", var_desc_->Name()),
var_->Get<framework::LoDTensor>().dims(), 0.0, scope);
}
AddTo(grad, grads_);
VLOG(3) << "grad_ after apply var grad " << var_desc_->Name() << " "
<< grads_->Get<framework::LoDTensor>().data<float>()[0];
}
std::vector<Variable*> OpBase::ApplyGrad(framework::Scope* scope) {
VLOG(3) << "op grad " << grad_op_desc_->Type();
for (const std::string& grad_invar : grad_op_desc_->InputArgumentNames()) {
if (grad_to_var_->find(grad_invar) == grad_to_var_->end()) {
// grad op inputs can be forward inputs, so not in grad_to_var.
continue;
}
VLOG(3) << "op grad in var " << grad_invar;
block_->FindRecursiveOrCreateVar(grad_invar);
framework::Variable* var = scope->Var(grad_invar);
const std::string& invar = grad_to_var_->at(grad_invar);
for (VarBase* varbase : *output_vars_) {
// Use the accumulated grads_ by sharing the input with grads_.
if (varbase->var_desc_->Name() == invar) {
var->GetMutable<framework::LoDTensor>()->ShareDataWith(
varbase->grads_->Get<framework::LoDTensor>());
break;
}
}
}
for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
VLOG(3) << "grad outvar " << outvar;
block_->FindRecursiveOrCreateVar(outvar);
framework::Variable* var = scope->Var(outvar);
if (!var->IsInitialized()) {
framework::VarDesc* var_desc = block_->FindVar(outvar);
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
var->GetMutable<framework::LoDTensor>();
} else {
LOG(ERROR) << "tracer doesn't support yet";
}
}
}
grad_op_desc_->InferShape(*block_);
grad_op_desc_->InferVarType(block_);
std::unique_ptr<framework::OperatorBase> opbase =
framework::OpRegistry::CreateOp(*grad_op_desc_);
opbase->Run(*scope, platform::CPUPlace());
// `ret` matches exactly with `input_vars_` of forward op.
std::vector<Variable*> ret;
for (size_t i = 0; i < input_vars_->size(); ++i) {
bool found = false;
for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
Variable* var = scope->FindVar(outvar);
VarBase* origin_var = (*input_vars_)[i];
std::string orig_var = grad_to_var_->at(outvar);
PADDLE_ENFORCE(origin_var->var_desc_->Name() == orig_var);
VLOG(3) << "apply grad " << outvar << " with origin " << orig_var;
origin_var->ApplyGrad(scope, var);
found = true;
ret.push_back(var);
// TODO(panyx0718): There might be another outvar with the same name.
// In that case, it doesn't matter the first one or the second one is
// used.
break;
}
if (!found) {
ret.push_back(nullptr);
}
}
return ret;
}
void VarBase::RunBackward(framework::Scope* scope) {
grads_ = CreateVariable(framework::GradVarName(var_desc_->Name()),
var_->Get<framework::LoDTensor>().dims(), 1.0, scope,
false);
if (!pre_op_) return;
Autograd(scope).RunBackward(this);
}
} // namespace imperative
} // namespace paddle
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace imperative {
class OpBase;
class VarBase {
public:
VarBase()
: pre_op_(nullptr),
pre_op_out_idx_(-1),
var_desc_(nullptr),
var_(nullptr),
grads_(nullptr) {}
virtual ~VarBase() {}
void ApplyGrad(framework::Scope* scope, framework::Variable* grad);
void RunBackward(framework::Scope* scope);
framework::LoDTensor& Grad();
OpBase* pre_op_;
int pre_op_out_idx_;
framework::VarDesc* var_desc_;
framework::Variable* var_;
framework::Variable* grads_;
};
class OpBase {
public:
OpBase()
: input_vars_(new std::vector<VarBase*>()),
output_vars_(new std::vector<VarBase*>()),
pre_ops_(new std::vector<OpBase*>()),
pre_ops_out_idx_(new std::vector<int>()),
op_desc_(nullptr),
grad_op_desc_(nullptr) {}
virtual ~OpBase() {
delete input_vars_;
delete output_vars_;
delete pre_ops_;
delete pre_ops_out_idx_;
if (grad_op_desc_) delete grad_op_desc_;
if (grad_to_var_) delete grad_to_var_;
}
std::vector<framework::Variable*> ApplyGrad(framework::Scope* scope);
std::vector<VarBase*>* input_vars_;
std::vector<VarBase*>* output_vars_;
std::vector<OpBase*>* pre_ops_;
std::vector<int>* pre_ops_out_idx_;
framework::OpDesc* op_desc_;
framework::OpDesc* grad_op_desc_;
std::unordered_map<std::string, std::string>* grad_to_var_;
framework::BlockDesc* block_;
};
class Layer {
public:
virtual ~Layer() {}
virtual std::vector<VarBase> Forward(const std::vector<VarBase>& inputs) {
std::vector<VarBase> vars;
return vars;
}
virtual void Backward() { LOG(ERROR) << "To support customize"; }
};
} // namespace imperative
} // namespace paddle
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/imperative/tracer.h"
namespace paddle {
namespace imperative {} // namespace imperative
} // namespace paddle
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/imperative/engine.h"
#include "paddle/fluid/imperative/layer.h"
namespace paddle {
namespace imperative {
void CreateGradOp(const framework::OpDesc& op_desc,
const std::unordered_set<std::string>& no_grad_set,
const std::vector<framework::BlockDesc*>& grad_sub_block,
framework::OpDesc** grad_op_desc,
std::unordered_map<std::string, std::string>* grad_to_var) {
std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
framework::OpInfoMap::Instance()
.Get(op_desc.Type())
.GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
// TODO(panyx0718): Leak?
*grad_op_desc = grad_op_descs[0].release();
}
class Tracer {
public:
explicit Tracer(framework::BlockDesc* root_block) : root_block_(root_block) {
root_scope_ = new framework::Scope();
scopes_[root_block_] = root_scope_;
}
virtual ~Tracer() { delete root_scope_; }
void Trace(OpBase* op, const std::vector<VarBase*>& inputs,
const std::vector<VarBase*>& outputs,
framework::BlockDesc* block) {
framework::Scope* scope = GetScope(block);
framework::OpDesc* op_desc = op->op_desc_;
VLOG(3) << "tracer tracing " << op_desc->Type();
op_desc->InferShape(*block);
op_desc->InferVarType(block);
std::unique_ptr<framework::OperatorBase> op_base =
framework::OpRegistry::CreateOp(*op_desc);
*op->input_vars_ = inputs;
for (VarBase* input : inputs) {
const std::string vname = input->var_desc_->Name();
framework::Variable* var = scope->Var(vname);
input->var_ = var;
if (!var->IsInitialized()) {
framework::VarDesc* var_desc = block->FindVar(vname);
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
var->GetMutable<framework::LoDTensor>();
} else {
LOG(ERROR) << "tracer doesn't support yet";
}
}
if (input->pre_op_) {
op->pre_ops_->push_back(input->pre_op_);
op->pre_ops_out_idx_->push_back(input->pre_op_out_idx_);
} else {
op->pre_ops_->push_back(nullptr);
}
}
*op->output_vars_ = outputs;
for (size_t i = 0; i < outputs.size(); ++i) {
const std::string vname = outputs[i]->var_desc_->Name();
framework::Variable* var = scope->Var(vname);
if (!var->IsInitialized()) {
framework::VarDesc* var_desc = block->FindVar(vname);
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
var->GetMutable<framework::LoDTensor>();
} else {
LOG(ERROR) << "tracer doesn't support yet";
}
}
outputs[i]->var_ = var;
outputs[i]->pre_op_ = op;
outputs[i]->pre_op_out_idx_ = i;
}
op_base->Run(*scope, platform::CPUPlace());
framework::OpDesc* grad_op_desc;
auto grad_to_var = new std::unordered_map<std::string, std::string>();
CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var);
op->grad_op_desc_ = grad_op_desc;
op->grad_to_var_ = grad_to_var;
op->block_ = block;
}
framework::Scope* GetScope(framework::BlockDesc* block) {
if (scopes_.find(block) != scopes_.end()) {
return scopes_.at(block);
}
framework::BlockDesc* parent_block = block->ParentBlock();
PADDLE_ENFORCE(scopes_.find(parent_block) != scopes_.end());
framework::Scope* scope = &scopes_[parent_block]->NewScope();
scopes_[block] = scope;
return scope;
}
private:
std::map<framework::BlockDesc*, framework::Scope*> scopes_;
framework::BlockDesc* root_block_;
framework::Scope* root_scope_;
};
} // namespace imperative
} // namespace paddle
set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune feed_fetch_method pass_builder parallel_executor profiler)
set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc)
set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune feed_fetch_method pass_builder parallel_executor profiler layer)
set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc imperative.cc)
if(WITH_PYTHON)
if(WITH_AMD_GPU)
hip_library(paddle_pybind SHARED
......
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/pybind/imperative.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/imperative/tracer.h"
namespace paddle {
namespace pybind {
// Bind Methods
void BindTracer(pybind11::module *m) {
pybind11::class_<imperative::Tracer>(*m, "Tracer", "")
.def("__init__",
[](imperative::Tracer &self, framework::BlockDesc *root_block) {
new (&self) imperative::Tracer(root_block);
})
.def("trace", &imperative::Tracer::Trace)
.def("get_scope", &imperative::Tracer::GetScope,
pybind11::return_value_policy::reference);
}
} // namespace pybind
} // namespace paddle
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <Python.h>
#include <vector>
#include "paddle/fluid/imperative/layer.h"
#include "pybind11/pybind11.h"
#include "pybind11/stl.h"
namespace paddle {
namespace pybind {
class PyLayer : public imperative::Layer {
public:
using imperative::Layer::Layer; // Inherit constructors
std::vector<imperative::VarBase> Forward(
const std::vector<imperative::VarBase>& inputs) override {
PYBIND11_OVERLOAD(std::vector<imperative::VarBase>, Layer, Forward,
inputs); // NOLINT
}
void Backward() override {
PYBIND11_OVERLOAD(void, Layer, Backward, ); // NOLINT
}
};
class PyOpBase : public imperative::OpBase {
public:
using imperative::OpBase::OpBase; // Inherit constructors
};
class PyVarBase : public imperative::VarBase {
public:
using imperative::VarBase::VarBase; // Inherit constructors
};
void BindTracer(pybind11::module* m);
} // namespace pybind
} // namespace paddle
......@@ -34,6 +34,7 @@ limitations under the License. */
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/framework/version.h"
#include "paddle/fluid/imperative/layer.h"
#include "paddle/fluid/memory/allocation/allocator_strategy.h"
#include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
......@@ -45,6 +46,7 @@ limitations under the License. */
#include "paddle/fluid/pybind/async_executor_py.h"
#include "paddle/fluid/pybind/const_value.h"
#include "paddle/fluid/pybind/exception.h"
#include "paddle/fluid/pybind/imperative.h"
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h" // NOLINT
#include "paddle/fluid/pybind/recordio.h"
......@@ -100,6 +102,42 @@ PYBIND11_MODULE(core, m) {
BindException(&m);
py::class_<imperative::VarBase, PyVarBase>(m, "VarBase", R"DOC()DOC")
.def(py::init<>())
.def("_run_backward",
[](imperative::VarBase &self, framework::Scope *scope) {
self.RunBackward(scope);
})
.def("_grad", &imperative::VarBase::Grad)
.def_property(
"desc",
[](const imperative::VarBase &self) { return self.var_desc_; },
[](imperative::VarBase &self, framework::VarDesc *var_desc) {
self.var_desc_ = var_desc;
},
py::return_value_policy::reference);
py::class_<imperative::OpBase, PyOpBase>(m, "OpBase", R"DOC()DOC")
.def(py::init<>())
.def_property(
"desc", [](const imperative::OpBase &self) { return self.op_desc_; },
[](imperative::OpBase &self, framework::OpDesc *op_desc) {
if (op_desc) {
self.op_desc_ = op_desc;
}
},
py::return_value_policy::reference);
py::class_<imperative::Layer, PyLayer /* <--- trampoline*/> layer(m, "Layer");
layer.def(py::init<>())
.def("forward",
[](imperative::Layer &self,
const std::vector<imperative::VarBase> &inputs) {
return self.Forward(inputs);
})
.def("backward", &imperative::Layer::Backward);
BindTracer(&m);
py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
.def_buffer(
[](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
......@@ -601,6 +639,7 @@ All parameter, weight, gradient are variables in Paddle.
m.def("set_feed_variable", framework::SetFeedVariable);
m.def("get_fetch_variable", framework::GetFetchVariable);
m.def("get_variable_tensor", framework::GetVariableTensor);
m.def("_is_program_version_supported", IsProgramVersionSupported);
......
......@@ -34,6 +34,7 @@ from . import io
from . import evaluator
from . import initializer
from . import layers
from . import imperative
from . import contrib
from . import nets
from . import optimizer
......@@ -67,6 +68,7 @@ __all__ = framework.__all__ + executor.__all__ + \
'initializer',
'layers',
'contrib',
'imperative',
'transpiler',
'nets',
'optimizer',
......
......@@ -18,6 +18,7 @@ import collections
import contextlib
import re
import six
import sys
import numpy as np
......@@ -49,6 +50,16 @@ GRAD_VAR_SUFFIX = core.kGradVarSuffix()
ZERO_VAR_SUFFIX = core.kZeroVarSuffix()
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()
_imperative_tracer_ = None
def _in_imperative_mode():
return _imperative_tracer_ is not None
def _imperative_tracer():
return _imperative_tracer_
class NameScope(object):
def __init__(self, name="", parent=None):
......@@ -202,7 +213,7 @@ def _debug_string_(proto, throw_on_error=True):
return proto.__str__()
class Variable(object):
class Variable(core.VarBase):
"""
In Fluid, every input and output of an operator is a variable. In most
cases, variables are used for holding different kinds of data or training
......@@ -266,6 +277,7 @@ class Variable(object):
stop_gradient=False,
is_data=False,
**kwargs):
core.VarBase.__init__(self)
self.block = block
self.error_clip = error_clip
......@@ -346,6 +358,18 @@ class Variable(object):
self.stop_gradient = stop_gradient
self.is_data = is_data
def _numpy(self):
scope = _imperative_tracer().get_scope(self.block.desc)
tensor = core.get_variable_tensor(scope, self.desc.name())
return np.array(tensor)
def _backward(self):
scope = _imperative_tracer().get_scope(self.block.desc)
self._run_backward(scope)
def _gradient(self):
return np.array(self._grad())
def __str__(self):
return self.to_string(True)
......@@ -492,7 +516,7 @@ class OpProtoHolder(object):
}
class Operator(object):
class Operator(core.OpBase):
"""
In Fluid, all the operation are represented by Operator, and Operator
is regarded as a build in an instruction of a Block. Users can use the
......@@ -548,6 +572,7 @@ class Operator(object):
inputs=None,
outputs=None,
attrs=None):
core.OpBase.__init__(self)
self.block = block
self.desc = desc
# note: not add self.attrs here:
......@@ -587,6 +612,7 @@ class Operator(object):
return True
return False
self.inputs = []
if inputs is not None:
for in_proto in proto.inputs:
found = find_name(inputs, in_proto.name)
......@@ -613,6 +639,13 @@ class Operator(object):
else:
self.desc.set_input(in_proto.name, [])
for inp in inputs.values():
if isinstance(inp, Variable):
self.inputs.append(inp)
elif isinstance(inp, list) or isinstance(inp, tuple):
self.inputs.extend(inp[:])
self.outputs = []
if outputs is not None:
given = set()
need = set()
......@@ -641,6 +674,12 @@ class Operator(object):
arg.op = self
self.desc.set_output(out_proto.name, out_arg_names)
for out in outputs.values():
if isinstance(out, Variable):
self.outputs.append(out)
elif isinstance(out, list) or isinstance(out, tuple):
self.outputs.extend(out[:])
if op_attrs is not None:
if not isinstance(op_attrs, dict):
raise TypeError("'attrs' should be a dict.")
......@@ -1206,6 +1245,8 @@ class Block(object):
"""
op_desc = self.desc.append_op()
op = Operator(block=self, desc=op_desc, *args, **kwargs)
if _in_imperative_mode():
_imperative_tracer().trace(op, op.inputs, op.outputs, self.desc)
self.ops.append(op)
return op
......@@ -2209,3 +2250,12 @@ def _get_var(name, program=None):
assert isinstance(program, Program)
return program.global_block().var(name)
@contextlib.contextmanager
def _imperative_guard(tracer):
global _imperative_tracer_
tmp_trace = _imperative_tracer_
_imperative_tracer_ = tracer
yield
_imperative_tracer_ = tmp_trace
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import print_function
from . import base
from .base import *
from . import layers
from .layers import *
__all__ = []
__all__ += layers.__all__
__all__ += base.__all__
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import contextlib
import numpy as np
from paddle.fluid import core
from paddle.fluid import framework
__all__ = ['enabled', 'guard', 'to_variable']
def enabled():
return framework._in_imperative_mode()
@contextlib.contextmanager
def guard():
train = framework.Program()
startup = framework.Program()
tracer = core.Tracer(train.current_block().desc)
with framework.program_guard(train, startup):
with framework.unique_name.guard():
with framework._imperative_guard(tracer):
yield
def to_variable(value, block=None):
if isinstance(value, np.ndarray):
if not block:
block = framework.default_main_program().current_block()
py_var = framework.Variable(
block,
type=core.VarDesc.VarType.LOD_TENSOR,
name=None,
shape=value.shape,
dtype=value.dtype)
scope = framework._imperative_tracer().get_scope(block.desc)
var = scope.var(py_var.name)
tensor = var.get_tensor()
tensor.set(value, core.CPUPlace())
return py_var
elif isinstance(value, framework.Variable):
return value
else:
raise ValueError("Unsupported type %s" % type(value))
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import contextlib
import sys
import numpy as np
from paddle.fluid import core
from paddle.fluid import framework
from paddle.fluid.imperative import base
__all__ = ['PyLayer']
class PyLayer(core.Layer):
def __init__(self):
pass
def __call__(self, inputs):
# TODO(panyx0718): Support declarative mode as well.
assert base.enabled()
if not isinstance(inputs, list) and not isinstance(inputs, tuple):
inputs = [inputs]
var_inputs = []
for x in inputs:
py_var = base.to_variable(x)
var_inputs.append(py_var)
outputs = self.forward(var_inputs)
return outputs
def forward(self, inputs):
return []
......@@ -17,10 +17,13 @@ from __future__ import print_function
import copy
import itertools
import six
import sys
import numpy as np
from .framework import Variable, Parameter, default_main_program, default_startup_program, dtype_is_floating
from . import unique_name
from paddle.fluid.initializer import Constant, Xavier
from paddle.fluid.imperative import base
from .param_attr import ParamAttr, WeightNormParamAttr
from . import core
from six.moves import zip
......@@ -46,23 +49,21 @@ class LayerHelper(object):
def startup_program(self):
return default_startup_program()
def to_variable(self, x):
return base.to_variable(x, self.main_program.current_block())
def append_op(self, *args, **kwargs):
return self.main_program.current_block().append_op(*args, **kwargs)
def multiple_input(self, input_param_name='input'):
inputs = self.kwargs.get(input_param_name, [])
type_error = TypeError(
"Input of {0} layer should be Variable or sequence of Variable".
format(self.layer_type))
if isinstance(inputs, Variable):
inputs = [inputs]
elif not isinstance(inputs, list) and not isinstance(inputs, tuple):
raise type_error
ret = []
if isinstance(inputs, list) or isinstance(inputs, tuple):
for inp in inputs:
ret.append(self.to_variable(inp))
else:
for each in inputs:
if not isinstance(each, Variable):
raise type_error
return inputs
ret.append(self.to_variable(inputs))
return ret
def input(self, input_param_name='input'):
inputs = self.multiple_input(input_param_name)
......
......@@ -6623,7 +6623,8 @@ def relu(x, name=None):
helper = LayerHelper('relu', **locals())
dtype = helper.input_dtype(input_param_name='x')
out = helper.create_variable_for_type_inference(dtype)
helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
helper.append_op(
type="relu", inputs={"X": helper.input('x')}, outputs={"Out": out})
return out
......
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import sys
import numpy as np
import paddle.fluid as fluid
from paddle.fluid import core
class MyLayer(fluid.imperative.PyLayer):
def __init__(self):
super(MyLayer, self).__init__()
def forward(self, inputs):
x = fluid.layers.relu(inputs[0])
self._x_for_debug = x
return [fluid.layers.elementwise_mul(x, x)]
class TestImperative(unittest.TestCase):
def test_layer(self):
with fluid.imperative.guard():
cl = core.Layer()
cl.forward([])
l = fluid.imperative.PyLayer()
l.forward([])
def test_layer_in_out(self):
with fluid.imperative.guard():
l = MyLayer()
x = l(np.array([1.0, 2.0, -1.0], dtype=np.float32))[0]
self.assertIsNotNone(x)
sys.stderr.write("%s output: %s\n" % (x, x._numpy()))
x._backward()
sys.stderr.write("grad %s\n" % l._x_for_debug._gradient())
if __name__ == '__main__':
unittest.main()
......@@ -101,6 +101,7 @@ packages=['paddle',
'paddle.dataset',
'paddle.reader',
'paddle.fluid',
'paddle.fluid.imperative',
'paddle.fluid.proto',
'paddle.fluid.proto.profiler',
'paddle.fluid.layers',
......
......@@ -27,6 +27,8 @@ import pydoc
member_dict = collections.OrderedDict()
experimental_namespace = {"paddle.fluid.imperative"}
def visit_member(parent_name, member):
cur_name = ".".join([parent_name, member.__name__])
......@@ -51,6 +53,8 @@ def visit_member(parent_name, member):
def visit_all_module(mod):
if (mod.__name__ in experimental_namespace):
return
for member_name in (
name
for name in (mod.__all__ if hasattr(mod, "__all__") else dir(mod))
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册